先进的计算机体系结构

CSCE930AdvancedComputer

Architecture

Introductions

Adoptedfrom

ProfessorDavidPatterson

&

DavidCuller

ElectricalEngineeringandComputerSciences

UniversityofCalifornia,Berkeley

Outline

•ComputerScienceataCrossroads:Parallelism

-Architecture:multi-coreandmany-cores

-Program:multi-threading

•ParallelArchitecture

-WhatisParallelArchitecture?

-WhyParallelArchitecture?

-EvolutionandConvergenceofParallelArchitectures

-FundamentalDesignIssues

•ParallelPrograms

-Whybotherwithprograms?

-Importantforwhom?

•Memory&StorageSubsystemArchitectures

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction2

Crossroads:ConventionalWisdominComp.Arch

•OldConventionalWisdom:Powerisfree,Transistorsexpensive

•NewConventionalWisdom:"Powerwall“Powerexpensive,Xtorsfree

(Canputmoreonchipthancanaffordtoturnon)

•OldCW:SufficientlyincreasingInstructionLevelParallelismvia

compilers,innovation(Out-of-order,speculation,VLIW,...)

•NewCW:"ILPwall”lawofdiminishingreturnsonmoreHWforILP

•OldCW:Multipliesareslow,Memoryaccessisfast

•NewCW:"Memorywall”Memoryslow,multipliesfast

(200clockcyclestoDRAMmemory,4clocksformultiply)

•OldCW:Uniprocessorperformance2X/1.5yrs

•NewCW:PowerWall+ILPWall+MemoryWall=BrickWall

-Uniprocessorperformancenow2XI5(?)yrs

=>Seachangeinchipdesign:multiple“cores”

(2Xprocessorsperchip/〜2years)

»Moresimplerprocessorsaremorepowerefficient

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction3

Crossroads:UniprocessorPerformance

10000

0(1000

Z8

/=

x，

v

>

w

>

8100

U

E

E

JO」

do」

10

1

197819801982198419861988199019921994199619982000200220042006

•VAX:25%/year1978to1986

•RISC+x86:52%/year1986to2002

•RISC+x86:??%/year2002topresent

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction4

SeaChangeinChipDesign

•Intel4004(1971):4-bitprocessor,

2312transistors,0.4MHz,

10micronPMOS,11mm2chip

・RISCII(1983):32-bit,5stage

pipeline,40,760transistors,3MHz,

3micronNMOS,60mm2chip

•125mm2chip,0.065micronCMOS

=2312RISCll+FPU+lcache+Dcache

-RISCIIshrinksto〜0.02mm2at65nm

-CachesviaDRAMor1transistorSRAM

-ProximityCommunicationviacapacitivecouplingat>1TB/s?

(IvanSutherland@SunIBerkeley)

•Processoristhenewtransistor?

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction5

Dejavualloveragain?

•Multiprocessorsimminentin1970s,’80s,'90s,…

•“・・・today'sprocessors...arenearinganimpasseas

technologiesapproachthespeedoflight..”

DavidMitchell,TheTransputer:TheTimeIsNow(1989)

•Transputerwaspremature

nCustommultiprocessorsstrovetoleaduniprocessors

nProcrastinationrewarded:2Xseq.perf./1.5years

•"Wearededicatingallofourfutureproductdevelopmentto

multicoredesignsThisisaseachangeincomputing”

PaulOtellini,President,Intel(2004)

•Differenceisallmicroprocessorcompaniesswitchto

multiprocessors(AMD,Intel,IBM,Sun;allnewApples2CPUs)

nProcrastinationpenalized:2Xsequentialperf.I5yrs

nBiggestprogrammingchallenge:1to2CPUs

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction6

ProblemswithSeaChange

Algorithms,ProgrammingLanguages,Compilers,

OperatingSystems,Architectures,Libraries,...not

readytosupplyThreadLevelParallelismorData

LevelParallelismfor1000CPUs/chip,

Architecturesnotreadyfor1000CPUs/chip

UnlikeInstructionLevelParallelism,cannotbesolvedbyjustby

computerarchitectsandcompilerwritersalone,butalsocannot

besolvedwithoutparticipationofcomputerarchitects

ThiscourseexploresISL(InstructionLevel

Parallelism)anditsshifttoThreadLevelParallelism

IDataLevelParallelism

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction7

Outline

•ComputerScienceataCrossroads:Parallelism

-Architecture:multi-coreandmany-cores

-Program:multi-threading

•ParallelArchitecture

-WhatisParallelArchitecture?

-WhyParallelArchitecture?

-EvolutionandConvergenceofParallelArchitectures

-FundamentalDesignIssues

•ParallelPrograms

-Whybotherwithprograms?

-Importantforwhom?

•Memory&StorageSubsystemArchitectures

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction8

WhatisParallelArchitecture?

•Aparallelcomputerisacollectionofprocessing

elementsthatcooperatetosolvelargeproblems

fast

•Somebroadissues:

-ResourceAllocation:

»howlargeacollection?

»howpowerfularetheelements?

»howmuchmemory?

-Dataaccess,CommunicationandSynchronization

»howdotheelementscooperateandcommunicate?

»howaredatatransmittedbetweenprocessors?

»whataretheabstractionsandprimitivesforcooperation?

-PerformanceandScalability

»howdoesitalltranslateintoperformance?

»howdoesitscale?

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction9

WhyStudyParallelArchitecture?

Roleofacomputerarchitect:

Todesignandengineerthevariouslevelsofacomputer

systemtomaximizeperformanceandprogrammability

withinlimitsoftechnologyandcost

Parallelism:

•Providesalternativetofasterclockforperformance

•Appliesatalllevelsofsystemdesign

•Isafascinatingperspectivefromwhichtoview

architecture

•Isincreasinglycentralininformationprocessing

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction10

WhyStudyitToday?

•History:diverseandinnovativeorganizational

structures,oftentiedtonovelprogrammingmodels

•Rapidlymaturingunderstrongtechnological

constraints

-The“killermicro“isubiquitous

-Laptopsandsupercomputersarefundamentallysimilar!

-Technologicaltrendscausediverseapproachestoconverge

•Technologicaltrendsmakeparallelcomputing

inevitable

-Inthemainstreamwiththerealityofmulti-coresandmany-cores

•Needtounderstandfundamentalprinciplesand

designtradeoffs,notjusttaxonomies

-Naming,Ordering,Replication,Communicationperformance

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction11

InevitabilityofParallelComputing

•Applicationdemands:Ourinsatiableneedforcomputing

cycles

-Scientificcomputing:VRsimulationsinBiology,Chemistry,Physics,...

-General-purposecomputing:Video,Graphics,CAD,Databases,AR,VI,

TP...

•TechnologyTrends

-Numberofcoresonchipgrowingrapidly(NewMoorsLaw)

-Clockratesexpectedtogouponlyslowly(tech,wall)

•ArchitectureTrends

-Instruction-levelparallelismvaluablebutlimited

-Coarser-levelparallelism,orthread-levelparallelism,themostviable

approach

•Economics

•Currenttrends:

一Today'smicroprocessorsaremultiprocessors

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction12

ApplicationTrends

•Demandforcyclesfuelsadvancesinhardware,andvice-

versa

-Cycledrivesexponentialincreaseinmicroprocessorperformance

-Drivesparallelarchitectureharder:mostdemandingapplications

•Rangeofperformancedemands

-Needrangeofsystemperformancewithprogressivelyincreasingcost

-Platformpyramid

•Goalofapplicationsinusingparallelmachines:Speedup

•Speedup(pprocessors)=&mrmance(pprocessors)一

Performance(1processor)

•Forafixedproblemsize(inputdataset),performance=

1/time

c.z.Time(1processor)

Speedupfixedproblem(pprocessors)=——-

11/5/2011CSCE930-AdvancedComputerArchitecture,IntroductiLff76(PProces^rs)

ScientificComputingDemand

GrandChallengeproblems

Globalchange

Humangenome

Fluidturbulence

LVehicledynamics

TBOceancirculation

Viscousfluiddynamics

Superconductormodeling

100GB一Quantumchromodynamics

Vision

①

E

①

--10GB-Structural

bn.biology

①Vehicle

」

①signaturePharmaceuticaldesign

1GB-

72'hour

weather

100MB-48-hour3Dplasma

weathermodelingChemicaldynamics

2DOilreservoir

10MB-

airfoilmodeling

Illi

WOMFLOPS1GFLOP510GFLOPS100GFLOPS1TFLOPS

Computationalperformancerequirement

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction14

EngineeringComputingDemand

•Largeparallelmachinesamainstayinmanyindustries

-Petroleum(reservoiranalysis)

-Automotive(crashsimulation,draganalysis,combustionefficiency),

-Aeronautics(airflowanalysis,engineefficiency,structuralmechanics,

electromagnetism),

-Computer-aideddesign

-Pharmaceuticals(molecularmodeling)

-Visualization

»Inalloftheabove

»Entertainment(3DfilmslikeAvatar&3Dgames)

»Architecture(walk-throughsandrendering)

»VirtualReality/lmmersion(museums,teleporting,etc)

-Financialmodeling(yieldandderivativeanalysis)

-Etc.

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction15

LearningCurveforParallelApplications

Numberofprocessors

•AMBERmoleculardynamicssimulationprogram

•StartingpointwasvectorcodeforCray-1

•145MFLOPonCray90,406forfinalversionon128-processor

Paragon,891on128-processorCrayT3D

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction16

CommercialComputing

•Alsoreliesonparallelismforhighend

-Scalenotsolarge,butusemuchmorewide-spread

-Computationalpowerdeterminesscaleofbusinessthatcanbehandled

•Databases,online-transactionprocessing,decision

support,datamining,datawarehousing...

•TPCbenchmarks(TPC-Corderentry,TPC-Ddecision

support)

-Explicitscalingcriteriaprovided

-Sizeofenterprisescaleswithsizeofsystem

-Problemsizenolongerfixedaspincreases,sothroughputisusedasa

performancemeasure(transactionsperminuteortpm)

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction17

SimilarStoryforStorage

•Divergencebetweenmemorycapacityandspeedmore

pronounced

-Capacityincreasedby1000xfrom1980-95,speedonly2x

-GigabitDRAMbyc.2000,butgapwithprocessorspeedmuchgreater

•Largermemoriesareslower,whileprocessorsget

faster

-Needtotransfermoredatainparallel

-Needdeepercachehierarchies

-Howtoorganizecaches?

•Parallelismincreaseseffectivesizeofeachlevelof

hierarchy,withoutincreasingaccesstime

•Parallelismandlocalitywithinmemorysystemstoo

-Newdesignsfetchmanybitswithinmemorychip;followwithfast

pipelinedtransferacrossnarrowerinterface

-Buffercachesmostrecentlyaccesseddata

•Diskstoo:Paralleldiskspluscaching

11/5/2011CSCE930-AdvancedComputerArchitecture,IrtTroduction18

Real-worldapplicationsdemandhigh-

performingandreliablestorage

HighperformanceComputingMedicinalImage

VirtualReality.lizailonnndImpingResenrchCentre

100TB100TBUniversityofHongKon«

Digitalbody

ITB/body

1PB5GB/day

11/5/2011

〔TheWorld|

PACHIC1//1V/M

OCEAN.一次ocri.v

诏可WA、

SOUTH

L.MLKICJ

\TLANTJC

PACIFIC2,060Mil

OCE4IV

GIS>1PBOceanresourcedat>1PB

Google,Yahoo,...

>lPB/year

Oilprospecting1PB

1PB=1000TB=1015Bytes,

Itisequaltothecapacityof10,000100GBdisks.

TechnoogyTrends:MoorecnLaw:2Xfrans一sfors/

飞

=yea飞JMXcores7nnyea

T

N

E

H

O

P

M

O

C

T/

S

O

C

G

IH

R

U

T

F^

U

N

A

M

E

IV

T

A

L6

E

R5

N4

S

TIO3

NT

EC2

NN—

EOU

HPF

TM0

OD3

FCE

OT8

2FA

GOR7

OG

LRE6

ET5

8IN

M4

UR

NE3

P2

-

•■・

IL,-■••-•,»・■卜

O

90123456789012345

56666666666777777

99999999999999999

II11111111111111

1

m>R

•^crammingMoreComponenfsonf。Wegrafedc-rcuifs:

IGordonMooreyErocfronicsy1965

•#onfrans一sfors/cosreffecHveinfegrafedcircuifdoubleeveryNmonfhs(12IANIA24)

11/5/2011CSCE930,AdvancedComputerArchifecture，ntroduc±on21

TrackingTechnologyPerformanceTrends

•Drilldowninto4technologies:

-Disks,

-Memory,

-Network,

-Processors

•Compare*1980Archaic(Nostalgic)vs.

*2000Modern(Newfangled)

-PerformanceMilestonesineachtechnology

•CompareforBandwidthvs.Latencyimprovements

inperformanceovertime

•Bandwidth:numberofeventsperunittime

-E.g.,Mbits/secondovernetwork,Mbytes/secondfromdisk

•Latency:elapsedtimeforasingleevent

-E.g.,one-waynetworkdelayinmicroseconds,

averagediskaccesstimeinmilliseconds

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction22

Disks:Archaic(Nostalgic)v.Modern(Newfangled)

CDCWrenI,1983Seagate373453,2003

3600RPM15000RPM(4X)

0.03GBytescapacity73.4GBytes(2500X)

Tracks/lnch:800Tracks/lnch:64000(80X)

Bits/lnch:9550Bits/lnch:533,000(60X)

Three5.25”plattersFour2.5,5platters

(in3.5”formfactor)

Bandwidth:Bandwidth:

0.6MBytes/sec86MBytes/sec(140X)

Latency:48.3msLatency:5.7ms(8X)

Cache:noneCache:8MBytes

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction23

LatencyLagsBandwidth(forlast〜20years)

PerformanceMilestones

Disk:3600,5400,7200,10000,

15000RPM(8x,143x)

(latency=simpleoperationw/ocontention

RelativeLatencyImprovementBW=best-case)

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction24

Memory:Archaic(Nostalgic)v.Modern(Newfangled)

•1980DRAM•2000DoubleDataRateSynchr.

(asynchronous)(clocked)DRAM

■0.06Mbits/chip•256.00Mbits/chip(4000X)

•64,000xtors,35mm2•256,000,000xtors,204mm2

•16-bitdatabusper•64-bitdatabusper

module,16pins/chipDIMM,66pins/chip(4X)

•13Mbytes/sec•1600Mbytes/sec(120X)

•Latency:225ns•Latency:52ns(4X)

•(noblocktransfer)•Blocktransfers(pagemode)

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction25

LatencyLagsBandwidth(last〜20years)

PerformanceMilestones

MemoryModule:16bitplain

DRAM,PageModeDRAM,32b,

64b,SDRAM,

DDRSDRAM(4x」20x)

Disk:3600,5400,7200,10000,

15000RPM(8x,143x)

(latency=simpleoperationw/ocontention

BW=best-case)

RelativeLatencyImprovement

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction26

LANs:Archaic(Nostalgic)v.Modern(Newfangled)

•Ethernet802.3•Ethernet802.3ae

•YearofStandard:1978•YearofStandard:2003

•10Mbits/s•10,000Mbits/s(1000X)

linkspeedlinkspeed

•Latency:3000|Lisec•Latency:190|Lisec(15X)

•Sharedmedia•Switchedmedia

•Coaxialcable•Category5copperwire

"Cat5"is4twistedpairsinbundle

CoaxialCable：/PlasticCoveringTwistedPair:

__________<,Braidedouterconductor

InsulatorXXZX二X

\J—CoppercoreCopper,1mmthick,

twistedtoavoidantennaeffect

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction27

LatencyLagsBandwidth(last〜20years)

•PerformanceMilestones

•Ethernet:10Mb,100Mb,

1000Mb,10000Mb/s(i6x,iooox)

•MemoryModule:16bitplain

DRAM,PageModeDRAM,32b,

64b,SDRAM,

DDRSDRAM(4x」20x)

•Disk:3600,5400,7200,10000,

15000RPM(8x,143x)

(latency=simpleoperationw/ocontention

RelativeLatencyImprovementBW=best-case)

11/5/2011CSCE930-AdvancedComputerArchitecture,Introduction28

CPUs:Archaic(Nostalgic)v.Modern(Newfangled)

1982Intel802862001IntelPentium4

12.5MHz